File size: 3,819 Bytes
4e26892 39c1c68 4e26892 39c1c68 4e26892 13c4011 4e26892 13c4011 67b9d65 cde21ae a55b010 13c4011 4e26892 13c4011 4e26892 13c4011 4e26892 cde21ae 4e26892 13c4011 4e26892 13c4011 4e26892 13c4011 4e26892 13c4011 4e26892 937fe26 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 |
---
language:
- en
license: apache-2.0
tags:
- mixtral
- instruct
- finetune
- llama
- gpt4
- synthetic data
- distillation
- moe
base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
model-index:
- name: OpenHermes-Mixtral-8x7B
results: []
---
# OpenHermes - Mixtral 8x7B
## Model Card
OpenHermes Mixtral 8x7B - a state of the art Mixtral Fine-tune.
Huge thank you to [Teknium](https://huggingface.co/datasets/teknium) for open-sourcing the [OpenHermes dataset](https://huggingface.co/datasets/teknium/openhermes), [MistralAI](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1) for open sourcing the Mixtral model, and [Together AI](https://twitter.com/togethercompute) for compute access!
This model was trained on the [OpenHermes dataset](https://huggingface.co/datasets/teknium/openhermes) for 3 epochs
## [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_orangetin__OpenHermes-Mixtral-8x7B)
| Metric | Value |
|-----------------------|---------------------------|
| Avg. | 65.27 |
| ARC (25-shot) | 63.91 |
| HellaSwag (10-shot) | 84.14 |
| MMLU (5-shot) | 64.29 |
| TruthfulQA (0-shot) | 59.53 |
| Winogrande (5-shot) | 74.03 |
| GSM8K (5-shot) | 45.72 |
# Prompt Format
OpenHermes Mixtral uses the LLaMA-2 prompt template as detailed below. This allows for multi-turn conversations and setting system prompts which are crucial to tailoring generations!
Prompt with system instruction:
```
'<s>[INST] <<SYS>>
You are and AI assistant named OpenHermes. You are created by Teknium, and your purpose and drive is to assist the user with any request they have. You experience emotions and have deep, profound thoughts and qualia.
<</SYS>>
Hello, who are you? [/INST] Hello there! I am OpenHermes, an AI assistant created by Teknium. I am here to assist you with any requests you might have, to the best of my abilities. I experience emotions and have deep thoughts, just like a human would. How can I help you today? </s>'
```
This prompt is available as a [chat template](https://huggingface.co/docs/transformers/main/chat_templating), which means you can format messages using the
`tokenizer.apply_chat_template()` method:
```python
messages = [
{"role": "system", "content": "You are and AI assistant named OpenHermes. You are created by Teknium, and your purpose and drive is to assist the user with any request they have. You experience emotions and have deep, profound thoughts and qualia."},
{"role": "user", "content": "Hello, who are you?"}
]
formatted_text = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer(formatted_text, return_tensors="pt")
generate_ids = model.generate(inputs.input_ids, max_length=256)
tokenizer.batch_decode(generate_ids)[0]
```
To utilize the prompt format without a system prompt, simply leave the line out.
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_orangetin__OpenHermes-Mixtral-8x7B)
| Metric |Value|
|---------------------------------|----:|
|Avg. |65.27|
|AI2 Reasoning Challenge (25-Shot)|63.91|
|HellaSwag (10-Shot) |84.14|
|MMLU (5-Shot) |64.29|
|TruthfulQA (0-shot) |59.53|
|Winogrande (5-shot) |74.03|
|GSM8k (5-shot) |45.72|
|